Stereoscopic Cameras Use AI to Get Smarter

Check out more Technology Talks videos.

Luxonis sells intelligent cameras that can implement machine-learning (ML) models directly instead of handing off raw video streams to a host (Fig. 1). Adding artificial intelligence (AI) to a camera offloads the host and reduces communication overhead. It also can simplify host software by providing processed data only when it's needed.

I talked with Luxonis' Bradley Dillon about the company's intelligent and stereoscopic cameras (see the video above).

1. Luxonis' OAK-1 is an intelligent, single camera (left) while the OAK-D is a stereoscopic camera with a regular color camera (right).

The OAK-1 and OAK-D are just two of the cameras in Luxonis' collection of products. The OAK-1 is a single 4K camera with a system that delivers 4 TOPS of compute power in the camera. The OAK-D provides similar performance, but it has two cameras for stereoscopic support as well as a central camera that provides video like the OAK-1 does on its own. Data from all three cameras is available for the computational subsystem, allowing for an ML model to combine all three when it's analyzing a video stream.

Intelligent cameras can address a range of applications without loading down the host with additional ML processing (Fig. 2).

2. Intelligent cameras can handle processing such as person detection, pose estimation, or smart dashcam applications with host processing.

Luxonis provides tools such as its DepthAI SDK, which supports the cameras and the creation of models to run on the systems. Most of the cameras employ a USB interface, though some are available with other interfaces such as Ethernet.